entailment temporal and aspectual and mark steedman thomas

29
Temporal and Aspectual Entailment Thomas Kober, Sander Bijl de Vroe and Mark Steedman University of Edinburgh

Upload: others

Post on 16-May-2022

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Entailment Temporal and Aspectual and Mark Steedman Thomas

Temporal and Aspectual EntailmentThomas Kober, Sander Bijl de Vroe and Mark Steedman

University of Edinburgh

Page 2: Entailment Temporal and Aspectual and Mark Steedman Thomas

Tense, Aspect & Entailment

Entailments become more involved with Tense...

Jane has arrived in London Jane is in London (now)

Jane will arrive in London Jane is in London (now)

Page 3: Entailment Temporal and Aspectual and Mark Steedman Thomas

Outline

● Tense, Aspect & Entailment

● Tense and Aspect with Distributional Models?

● Lets have some TEA: A new Entailment dataset

● Temporal Entailment with Distributional Models?

● Conclusion

Page 4: Entailment Temporal and Aspectual and Mark Steedman Thomas

Outline

● Tense, Aspect & Entailment

● Tense and Aspect with Distributional Models?

● Lets have some TEA: A new Entailment dataset

● Temporal Entailment with Distributional Models?

● Conclusion

Page 5: Entailment Temporal and Aspectual and Mark Steedman Thomas

Tense, Aspect & Entailment

In entailment rule mining, entailment typically applies to verb lemmas

Arrive in <LOC> be in <LOC>

Visit <LOC> arrive in <LOC>

Visit <LOC> leave <LOC>

Tense and Aspect typically ignored

(e.g. Szpektor and Dagan, 2008,Berant et al., 2011…)

Page 6: Entailment Temporal and Aspectual and Mark Steedman Thomas

Tense, Aspect & Entailment

However, Entailment heavily interacts with Tense & Aspect (and also Mood)

Jane has sold the house Jane owned the house

Jane has sold the house Jane will own the house

Jane has bought the house Jane owns the house

Jane has bought the house Jane owned the house

Page 7: Entailment Temporal and Aspectual and Mark Steedman Thomas

Tense, Aspect & Entailment

Consequences of the present perfect hold in the present

Elizabeth has gone to Meryton Elizabeth is in Meryton

Elizabeth went to Meryton Elizabeth is in Meryton

Elizabeth went to Meryton Elizabeth was in Meryton

Page 8: Entailment Temporal and Aspectual and Mark Steedman Thomas

Tense, Aspect & Entailment

Aspect also gives rise to the Imperfective Paradox (Dowty, 1986)

Mary was walking in the woods Mary walked in the woods

Jane was reaching London Jane reached London

(resolved through lexical aspect types and type coercion, Moens and Steedman (1988))

Page 9: Entailment Temporal and Aspectual and Mark Steedman Thomas

Outline

● Tense, Aspect & Entailment

● Tense and Aspect with Distributional Models?

● Lets have some TEA: A new Entailment dataset

● Temporal Entailment with Distributional Models?

● Conclusion

Page 10: Entailment Temporal and Aspectual and Mark Steedman Thomas

Tense and Aspect with Distributional Models?

Test whether morphosyntactic tense is encoded in vector spaces

Two Experiments

Auxiliary-Verb Agreement: Can we detect whether an inflected verb is combined with a correct or incorrect auxiliary? (inspired by Linzen et al. (2016))

Translation Operation: Can we learn a translation operation between different tenses? (inspired by Bolukbasi et al. (2016) and Shoemark et al. (2017))

Page 11: Entailment Temporal and Aspectual and Mark Steedman Thomas

5 models:Word2vec (Mikolov et al., 2013) APTs (Weir et al., 2016)fastText (Mikolov et al., 2018) ELMo (Peters et al., 2018) BERT (Devlin et al., 2018)

Pointwise addition as composition function for word2vec & fastTextDependency based composition function for APTsELMo, BERT, encoder for phrases and sentences

Tense and Aspect with Distributional Models?

Page 12: Entailment Temporal and Aspectual and Mark Steedman Thomas

Create representations for verb-auxiliary pairs, such as will visit, *will visitingis arriving, *has arriving

Simple binary classification task (correct combination vs. incorrect combination)

Extract verbs from OBWB, generate pairs, post-processing -> 36k aux-verb pairs,

Train logistic regression classifier. If it distinguishes correct from incorrect combinations, the embeddings encode some agreement information

Tense and Aspect with Distributional Models?Auxiliary-Verb Agreement

Page 13: Entailment Temporal and Aspectual and Mark Steedman Thomas

Tense and Aspect with Distributional Models?Auxiliary-Verb Agreement

Model Averaged Accuracy

Majority class 0.61

Word2vec 0.71

APTs 0.92

fastText 0.71

ELMo 0.59

BERT 0.91

Page 14: Entailment Temporal and Aspectual and Mark Steedman Thomas

All models decent except ELMo:Worse than majority class baseline. We also tried encoding full sentences instead of just auxiliary-verb phrases, but weren't able to improve performance

Word2vec and fastText: some combinations work better than others, but overall substantially weaker than BERT and APTs

BERT and APTs: very strong, APTs in particular benefit from sparse representation

Tense and Aspect with Distributional Models?Auxiliary-Verb Agreement

Page 15: Entailment Temporal and Aspectual and Mark Steedman Thomas

Averaged offset: Learn offset vector for each tense:Randomly choose 10 verbs; subtract lemma vector from inflected form vector; average the differences to obtain offset vector. Prediction: add offset vector to unseen lemma to predict inflected verb vectorTranslation matrix:2.3k verbs per tense from OBWB, learn a neural network for each tense that given lemma vector predicts the inflected form vector Evaluate: using MRR, the rank of the correct inflected verb in the neighbour list of the predicted verb

Tense and Aspect with Distributional Models? Translation Operation

Page 16: Entailment Temporal and Aspectual and Mark Steedman Thomas

Tense and Aspect with Distributional Models? Translation Operation

Results

ELMo BERT have contextualised embeddings, so they are at a disadvantage in comparison to word2vec & fasttext who have static embeddings. APTs also have static embeddings, but suffer from sparsity

Page 17: Entailment Temporal and Aspectual and Mark Steedman Thomas

Actually works fairly wellMost models don’t excel on both tasks, but morphosyntactic relations seem to be somehow encoded in each space Aux-Verb Agreement Translation Operation

w2v ✔

fastText ✔ ✔

ELMo ✔

BERT ✔

APT ✔

Tense and Aspect with Distributional Models?

Page 18: Entailment Temporal and Aspectual and Mark Steedman Thomas

Outline

● Tense, Aspect & Entailment

● Tense and Aspect with Distributional Models?

● Lets have some TEA: A new Entailment dataset

● Temporal Entailment with Distributional Models?

● Conclusion

Page 19: Entailment Temporal and Aspectual and Mark Steedman Thomas

TEA: A new Entailment dataset

There are a number of recently released large scale entailment datasets, e.g. SNLI (Bowman et al., 2015; Williams et al., 2017; Poliak et al., 2018; etc)

However, neither focuses on or includes a substantial proportion of examples focusing on Temporal Entailment

(Very limited number in e.g. FraCas, but nothing of larger scale)

Therefore, we created TEA (Temporal Entailment Assessment)

Available from here: https://github.com/tttthomasssss/iwcs2019

Page 20: Entailment Temporal and Aspectual and Mark Steedman Thomas

TEA contains 11,138 premise-hypothesis sentence pairs

The only differences between premise and hypothesis are the

Main verb

Morphosyntactic tense of the main verb

(a preposition or particle to make the sentence grammatical)

Labelled by 2 human annotators (Thomas & Sander)

TEA: A new Entailment dataset

Page 21: Entailment Temporal and Aspectual and Mark Steedman Thomas

Candidate entailment pairs were mined from previous entailment datasets and other lexical resources such as the WordNet (Fellbaum, 1998) verb entailment graph or the before-after category of VerbOcean (Chlovski and Pantel, 2004)

Arguments were manually added to form full sentences

The arguments also resolved some ambiguity

We tried to avoid habitual readings

All sentences follow the same syntactic structure

TEA: A new Entailment dataset

Page 22: Entailment Temporal and Aspectual and Mark Steedman Thomas

Examples:John is visiting London John has arrived in London 1John is visiting London John will leave London 1John is visiting London John has left London 0George is acquiring the house George owns the house 0

TEA: A new Entailment dataset

Page 23: Entailment Temporal and Aspectual and Mark Steedman Thomas

Outline

● Tense, Aspect & Entailment

● Tense and Aspect with Distributional Models?

● Lets have some TEA: A new Entailment dataset

● Temporal Entailment with Distributional Models?

● Conclusion

Page 24: Entailment Temporal and Aspectual and Mark Steedman Thomas

Evaluation:

Binary labels (entailment : no-entailment) with a class distribution 22 : 78

Average Precision (a.k.a. Area under the Precision/Recall curve), Accuracy & F1-Score

a) Run word2vec, APTs, fastText, ELMo and BERTb) Run biLSTMs pre-trained on SNLI or DNCc) Compare to two simple baselines

i) Majority classii) Majority class per Tense pair

Temporal Entailment with Distributional Models?

Page 25: Entailment Temporal and Aspectual and Mark Steedman Thomas

Model Avg. Precision Model Avg. Precision Accuracy F1-Score

word2vec 0.31 biLSTM-DNC 0.22 0.58 0.49

APT 0.28 biLSTM-SNLI 0.21 0.51 0.47

fastText 0.30 Majority class 0.22 0.78 0.44

ELMo 0.21 Maj. Class / Tense pair 0.35 0.80 0.66

BERT 0.27

Temporal Entailment with Distributional Models?

Page 26: Entailment Temporal and Aspectual and Mark Steedman Thomas

No model beats the majority class / tense pair baseline

All models achieve low scores - despite the simplicity of the sentences

Temporal Entailment is a difficult problem

Temporal Entailment with Distributional Models?

Page 27: Entailment Temporal and Aspectual and Mark Steedman Thomas

Outline

● Tense, Aspect & Entailment

● Tense and Aspect with Distributional Models?

● Lets have some TEA: A new Entailment dataset

● Temporal Entailment with Distributional Models?

● Conclusion

Page 28: Entailment Temporal and Aspectual and Mark Steedman Thomas

Conclusion

Highlighted the importance of Tense & Aspect for Natural Language Inference

Showed that the morphosyntax of Tense & Aspect is encoded in vector space (to a reasonable degree at least)

Introduced TEA - a dataset for Temporal Entailment (https://github.com/tttthomasssss/iwcs2019/)

But model performance is bad across the board on a semantic task

Distributional model governed by contextual similarity, not ideal for doing inference that depend on Tense & Aspect